keksuccino/better-codex-app-custom-provider-support

This project is an unofficial patch for the macOS ChatGPT and Codex desktop app that lets you pick a different AI model provider for each new task, without having to sign out of your regular ChatGPT account. Normally the desktop app only talks to OpenAI's own models through your logged in account, but this patch adds a provider menu on top of that, so you can send a task to something like OpenRouter instead while still keeping your OpenAI login active for tasks that should stay on ChatGPT's own models. The patch works by modifying the installed app directly. A Python installer script closes the running app, makes a full backup of it first, edits the app's packaged code, updates the integrity metadata Electron apps use to verify their own files have not been tampered with, and then re-signs the app so macOS will still run it. Because it edits the actual installed application rather than replacing it, the author warns that any future ChatGPT update will overwrite the patch, so the installer needs to be run again after each update, and that the tool stops itself rather than risk breaking the app if an update changes the code it expects to find. Setting up a custom provider means adding it to Codex's own configuration file, giving it a name, an API endpoint, and a way to read your API key, ideally from the Mac Keychain rather than typed directly into a config file. You can also add extra models that are not part of the default OpenAI lineup by exporting the built in model list, copying an entry with similar capabilities, and adjusting its name and details for the new model. A separate JSON file that the installer creates controls exactly which providers show up in the desktop menu and which specific models are automatically routed to which provider. One clear limitation is that once a conversation has started, the provider used for that conversation cannot be changed. The project only supports macOS for now, and the author is explicit that this is an unofficial, unsupported modification, used entirely at the user's own risk, with the possibility of losing chat history or running into problems with the ChatGPT account if something goes wrong.